# PINN drift surrogates: pass a 20% held-out MIDR test first

Ashley Coleman · November 1, 2026

> Learn how PINN drift surrogates meet 20% MIDR test criteria, with four key takeaways and a physics‑verified adoption rule for seismic demand prediction.

This guide provides a concise decision framework for using PINN drift surrogates in seismic demand prediction.

It lists four specific takeaways and a clear adoption rule based on error reduction and physics verification.

![This guide provides a concise decision framework for — PINN drift surrogates](https://static.mm-ais.com/article-images-ai/pinn-drift-surrogates-pass-a-20-held-out-ai-0adc155c.jpg)

## How a physics-informed drift surrogate works

A physics-informed drift surrogate trains a neural network with a composite objective rather than a single data-misfit term. One part measures error against simulated drift time histories; the other is a residual that penalizes violations of the governing dynamic equilibrium of the structural model. Because the residual can be evaluated where no labeled drift history exists, the network is discouraged from producing mechanically inadmissible predictions simply because they fit sparse training data.

| Loss component | What it penalizes | Where it can be evaluated |
| --- | --- | --- |
| Data misfit term | Difference between predicted and simulated drift time histories | Training points with labels |
| Equilibrium residual | Violation of governing dynamic equilibrium | Any collocation point, including unlabeled points |
| GIDS consistency | Departure from the spectrum-based MIDR estimate | Ground-motion descriptor space |

The physics anchor for that residual is the generalized interstory drift spectrum (GIDS), which approximates peak interstory drift ratio (MIDR) from ground-motion characteristics. The residual is built to stay consistent with this spectrum, so drift is not learned from data alone; it is constrained by a spectrum-based estimate of demand and by the equilibrium equations of the model. GIDS supplies the demand-side reference, while equilibrium supplies the structural-side constraint.

This split addresses two different failure modes. A data-only term can be unreliable where the training set is thin or where a ground motion lies outside the sampled distribution; an equilibrium or spectrum residual can be misleading where its underlying approximation is poor. The residual must therefore be treated as an active constraint, not as a decorative penalty, and its behavior should be monitored during training.

A practical check is to watch whether both loss components decrease together. If the data term falls while the residual plateaus, the optimizer is effectively ignoring physics and the model is reverting to a data-only surrogate. The GIDS anchor and the equilibrium residual should also use consistent ground-motion descriptors; otherwise the two constraints can pull the prediction in different directions. When both terms remain active and mutually consistent, the surrogate is genuinely physics-informed.

![How a physics-informed drift surrogate works — PINN drift surrogates](https://static.mm-ais.com/article-images-ai/pinn-drift-surrogates-pass-a-20-held-out-ai-b4468ec3.jpg)

## The evidence: where the 20% comes

The 20% MIDR reduction claim is best read as a target to validate locally, not as a figure already demonstrated in the sources cited here. A hybrid machine-learning study on seismic demand prediction (published via Springer) blends base learners for base shear and stacks learners for interstory drift ratio, delivering high-resolution, physically coherent demand predictions. That drift-specific stacking result is the closest published benchmark to the 20% claim, but the cited source text does not state a held-out MIDR error reduction percentage against a purely data-driven surrogate, and no author, volume, or page attribution for such a figure appears in the available sources.

The arXiv 2023 physics-informed time-domain simulation work supports a different part of the case: it reports that PINNs are "drastically accelerating computations of non-linear dynamical systems." That is a computational-speed claim for nonlinear dynamics, not a structural drift-accuracy benchmark, and it does not report MIDR error or a physics-residual verification against a nonlinear time-history reference.

| Source | Reported result | Relevance to the 20% MIDR claim |
| --- | --- | --- |
| Kaveh, Asian Journal of Civil Engineering, 2026, vol. 27, pp. 4807–4832 | Blended base learners for base shear; stacked learners for interstory drift ratio; "high-resolution, physically coherent" demand predictions | Closest published benchmark for drift-specific surrogate accuracy; no held-out MIDR percentage stated in the cited text |
| arXiv, 2023, physics-informed time-domain simulation | PINNs "drastically accelerating computations of non-linear dynamical systems" | Supports computational acceleration for nonlinear dynamics; no MIDR error or residual verification reported |

Taken together, the independent sources establish that hybrid stacking can produce physically coherent drift demands and that PINNs can accelerate nonlinear time-domain simulation. They do not, in the cited texts, jointly supply the held-out MIDR comparison or the residual check required by the adoption rule. The practical reading is therefore to treat the roughly 20% reduction as a hypothesis to be confirmed on the project's own held-out data and reference simulations.

## Options compared: PINN vs data-only vs spectrum

For peak interstory drift ratio prediction, the practical choice is among three surrogates. The decision should rest on held-out error and residual verification, not on training convenience alone, because a fast model that fails either check can misrank drift demand.

Option A, a data-only machine-learning surrogate, is fast to train and evaluate, but it carries no physics residual. Therefore, drift extrapolation outside the training ground-motion set is unconstrained, and prediction error grows where the data are sparse. Use it for screening or in regions of the ground-motion space already well represented by training data.

Option B, a physics-informed neural network surrogate, adds the equilibrium/GIDS residual. It wins only when held-out MIDR error drops by the required margin versus Option A and that residual is verified against nonlinear time-history reference runs. Both conditions are necessary; a residual that is not checked against reference runs does not justify adoption.

Option C, a conventional generalized interstory drift spectrum estimate, requires no training and remains transparent. It is the defensible fallback when the PINN does not satisfy both checks, and it is especially useful when auditability or a quick first approximation matters.

| Option | Principal strength | Principal limitation | Adopt when |
| --- | --- | --- | --- |
| A: data-only ML surrogate | Fast training and evaluation | No physics residual; unconstrained extrapolation | Screening or dense-data regions only |
| B: PINN surrogate | Physics residual plus data fit | Needs held-out error gain and residual verification | Both checks pass |
| C: conventional GIDS estimate | No training; transparent | Approximate first estimate | PINN checks fail or auditability matters |

Under the stated condition, Option B is the winner; if either the error gain or the residual verification fails, Option C is the defensible fallback, while Option A remains a fast screening surrogate. This rule keeps the surrogate choice tied to evidence rather than to model complexity.

## Costs and numbers that matter

Training cost is dominated by generating the physics-consistent training set. Each labeled sample requires a nonlinear time-history analysis, so the surrogate does not save compute on day one; its value is realized only after that generation cost is amortized across many predictions. Before adopting, estimate the number of future predictions needed to recover the generation cost, and compare that break-even volume with the expected project pipeline. For a small portfolio or a one-off study, the incumbent analysis workflow may remain cheaper.

The 20% margin is a relative reduction in held-out MIDR error, not an absolute drift value. If a data-only baseline has 15% mean absolute percentage error, the PINN must reach 12% or lower, because 15% × 0.80 = 12%. The same arithmetic applies to your own baseline: multiply the baseline error by 0.80 and require the PINN to beat that value. Compute the threshold on your own data and held-out split; do not assume that a published percentage transfers to your building set, ground-motion set, or error metric.

| Baseline held-out MIDR error | Required PINN error for a 20% reduction |
| --- | --- |
| 15% | ≤12% |
| Your measured baseline | ≤0.80 × baseline |

Verification carries its own cost. The physics residual must be verified against a nonlinear time-history reference, and each verification case is another nonlinear run. Budget a representative set of cases rather than a single spot check, and repeat verification when the model architecture, training data, ground-motion set, or preprocessing changes. Treat verification as part of the adoption cost, not as a post-hoc formality.

The decision is therefore a cost-and-error tradeoff. Add generation, training, and verification costs, then compare them with the cost of continuing to run nonlinear analyses directly. Adopt the PINN only when it clears both gates: at least a 20% reduction in held-out MIDR error and a verified physics residual. If either gate fails, retain the incumbent surrogate and revisit the decision when data volume or verification coverage improves.

Keep an auditable record with the model: baseline error, held-out protocol, verification cases, and cost per prediction. That record separates a reproducible cost advantage from a one-time benchmark and keeps the 20% claim tied to the conditions under which it was measured.

## What the evidence does NOT establish

The 20% MIDR error reduction is a benchmark threshold for adoption, not a result that any grounded source attributes to a PINN. The hybrid surrogate framework reported in the Springer-published seismic demand study is described as physically coherent, but not as delivering a quantified 20% improvement over a data-only baseline. Treat the 20% as a local acceptance criterion to be demonstrated on held-out data, not as an established performance margin.

The claim also weakens when the structure or loading lies outside the training distribution. Unusual hysteretic models, near-fault ground motions, and tall flexible frames can expose gaps that a physics residual alone may not close, because the residual enforces consistency with the modeled equations rather than supplying missing response data. Verify performance separately for each structural class and ground-motion set before adopting the surrogate.

Even the spectrum-based physics anchor has limits. The generalized interstory drift spectrum is useful as a first approximation, but differences between that spectrum and true responses mean its adjustment or correction remains necessary; it should not be treated as an exact substitute for nonlinear time-history verification. A residual checked only against the spectrum, rather than against a nonlinear time-history reference, does not establish the required margin.

| Limit | Why the 20% claim may fail | Required check |
| --- | --- | --- |
| No grounded source reports 20% for PINNs | Kaveh (2026) reports qualitative physical coherence, not a quantified 20% versus a data-only baseline | Demonstrate at least 20% held-out MIDR error improvement locally |
| Outside the training distribution | The residual cannot compensate for absent data for unusual hysteretics, near-fault motions, or tall flexible frames | Verify per structural class and ground-motion set |
| The spectrum is approximate | The generalized interstory drift spectrum differs from true responses and needs adjustment | Verify the residual against a nonlinear time-history reference |

In practice, the safe reading is narrower than the headline: the surrogate is justified only where held-out MIDR error improves by at least 20% and the residual is verified against nonlinear time-history results. Where either condition fails, the evidence does not support replacing the data-only baseline or the spectrum-based estimate.

## Verifying the 20% margin

To put the adoption rule into practice, run this concrete numeric walkthrough on a four-story shear-beam model (per the ASCE Library drift-estimation protocol) with a held-out ground-motion set. For each run, you must have both a data-only baseline and a PINN with the equilibrium/GIDS residual trained on the same split. The first action is to freeze both models, then test only on the held-out set. Record **Checkpoint 1** as the baseline MIDR error on that held-out set before looking at any PINN result. This strict order is the only way to prevent the margin from being inflated by post-hoc comparisons.

With both models fixed to the exact same held-out ground-motion set, compute **Checkpoint 2** as the PINN MIDR error. The pass condition is strictly arithmetic: the rule passes only if **Checkpoint 2 ≤ 0.8 × Checkpoint 1**. To see the check in action, treat the product as a hard ceiling for the PINN error. If the PINN error lands at or below that ceiling, the ≥20% cut is met by ratio alone. If it lands above that ceiling, the ratio fails no matter how smooth the training curves look. Do not swap in a different motion subset or a different MIDR error definition to force this product.

Ratio is not enough on its own. Compute **Checkpoint 3** as the physics residual magnitude of the PINN versus the nonlinear time-history reference for those same held-out cases, and judge it against a pre-agreed tolerance for this model class. The adoption rule requires both at the same time: the 0.8× ratio and Checkpoint 3 within tolerance. A passing ratio with a residual outside of tolerance must be rejected, and a residual inside of tolerance with a failing ratio must be rejected. This pairing is what keeps the claimed margin tied to enforced physics instead of data fit alone.

The three checkpoints above can be applied directly in a review: record the baseline MIDR error first, then the PINN error against the 0.8× ceiling, then the residual magnitude against the pre-agreed nonlinear time-history tolerance, and adopt only if both the ratio and the residual check pass.

## Decision rules for adopting a PINN surrogate

Adopt a physics-informed neural network (PINN) drift surrogate for a specific structural class and intensity range only when both of the following hold on held-out data. The PINN’s peak MIDR error must be no more than 80 percent of the data-only baseline’s peak MIDR error—equivalent to at least a 20 percent reduction—and the physics residual must pass verification against a nonlinear time-history reference. If either test fails, the PINN does not clear the adoption threshold.

When the margin is not demonstrated, retain the conventional generalized interstory drift spectrum (GIDS) estimate as the controlling design check and treat the PINN as exploratory only. The 20 percent figure is a local validation target, not a guaranteed transfer; as noted above, no cited source reports a quantified held-out MIDR improvement for a PINN, so hybrid surrogate performance must be confirmed on the project’s own held-out set rather than assumed from published results.

Distribution determines applicability. If the target structure or ground-motion set falls outside the PINN’s training distribution, re-run the verification checklist before any adoption decision. Clearing the threshold for one structural class or intensity range does not convey clearance to another, and a new ground-motion suite or retrained model invalidates prior clearance.

| Condition | Decision | Scope |
| --- | --- | --- |
| Held-out MIDR error ≤ 80% of baseline AND physics residual passes nonlinear reference check | Adopt for design use | Verified structural class and intensity range only |
| Margin not demonstrated | Keep GIDS as design check; PINN exploratory only | All cases |
| Target outside training distribution | Re-run verification checklist before adoption | Before any decision |

Document the held-out set, the baseline error, the PINN error, and the residual check results. Without that documentation, the conventional GIDS estimate remains the default design check, and the PINN should not be used to govern member sizing or drift limits.

Re-verification is also required when the structural class changes or when the intensity range shifts beyond the verified band. The adoption decision is tied to the model version, the ground-motion set, and the structural typology, not to the PINN architecture alone.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Run the held-out MIDR error comparison between the PINN drift surrogate and the data-only baseline. | Adoption requires beating the data-only baseline by ≥20% on held-out MIDR error. |
| 2 | Confirm the physics residual is embedded in the PINN loss function rather than applied after training. | The ≈20% MIDR error reduction is conditional on enforcing the physics residual in the loss. |
| 3 | Verify the PINN physics residual against a nonlinear time-history reference. | Verification against a nonlinear time-history reference is required alongside the error threshold. |
| 4 | Adopt the PINN drift surrogate for peak interstory drift ratio (MIDR) prediction only if both checks pass. | The decision rule mandates both the ≥20% error improvement and residual verification before adoption. |
| 5 | Keep the conventional drift-spectrum estimate as the design check if either check fails. | The surrogate is not adopted unless it clears both the error threshold and physics verification. |
| 6 | Record the held-out MIDR error and nonlinear time-history verification results in the design file. | Documents the basis for adopting or rejecting the PINN drift surrogate. |

## Frequently Asked Questions

**Where can the equilibrium residual be evaluated in a PINN drift surrogate?**

It can be evaluated at any collocation point, including unlabeled points.

**What does the data misfit term compare in the composite loss?**

It compares predicted and simulated drift time histories at training points with labels.

**What does GIDS consistency measure?**

It measures departure from the spectrum-based MIDR estimate in ground-motion descriptor space.

**What physical quantity does the generalized interstory drift spectrum (GIDS) approximate?**

It approximates peak interstory drift ratio (MIDR) from ground-motion characteristics.

**What does the equilibrium residual penalize?**

It penalizes violations of the governing dynamic equilibrium of the structural model.

**Why does the residual help avoid mechanically inadmissible predictions?**

Because the residual can be evaluated where no labeled drift history exists, the network is discouraged from producing mechanically inadmissible predictions simply because they fit sparse training data.

## Quick answers

| What does a physics-informed drift surrogate train a neural network with instead of a single data-misfit term? | It trains a neural network with a composite objective rather than a single data-misfit term. |
| --- | --- |
| What does the equilibrium residual penalize in the structural model? | It penalizes violations of the governing dynamic equilibrium of the structural model. |
| Where can the equilibrium residual be evaluated? | It can be evaluated at any collocation point, including unlabeled points. |
| What does the generalized interstory drift spectrum (GIDS) approximate? | It approximates peak interstory drift ratio (MIDR) from ground-motion characteristics. |
| What is the clear adoption rule for PINN drift surrogates based on? | It is based on error reduction and physics verification. |

Also worth reading: **Neural networks vs finite elements: 10,000 points—not a drift test**: [Neural networks vs finite elements:](https://aistructuralreview.com/blog/neural-networks-vs-finite-elements-10000-pointsnot-a-drift-test.php) · **OpenSees vs ABAQUS: 40% Convergence Gain in 2026 PBSD Frames**: [OpenSees vs ABAQUS: 40% Convergence](https://aistructuralreview.com/blog/opensees-vs-abaqus-40-convergence-gain-in-2026-pbsd-frames.php) · **OpenSees 41% Faster: 44-Record Benchmark With Receipts & Caveats**: [OpenSees 41% Faster: 44-Record Benchmark](https://aistructuralreview.com/blog/opensees-41-faster-44-record-benchmark-with-receipts-caveats.php)

### Related reading

- [Mesh-First Protocol vs Material Laws in Collapse Drift (2026)](https://aistructuralreview.com/blog/mesh-first-protocol-vs-material-laws-in-collapse-drift-2026.php)
- [Neural networks vs finite elements: 10,000 points—not a drift test](https://aistructuralreview.com/blog/neural-networks-vs-finite-elements-10000-pointsnot-a-drift-test.php)
- [Stronger Thigh Bones: Evolutionary Structural Optimization (ESO) — 2% Removal or Not?](https://aistructuralreview.com/blog/stronger-thigh-bones-evolutionary-structural-optimization-eso-2-removal-or-not.php)
- [Memphis building codes: 2018 International Building Code lateral retrofit vs rebuild](https://aistructuralreview.com/blog/memphis-building-codes-2018-international-building-code-lateral-retrofit-vs-rebuild.php)
- [2% Drift Limit: Verified Gate, Not Default, for 8-Story Steel Frames](https://aistructuralreview.com/blog/2-drift-limit-verified-gate-not-default-for-8-story-steel-frames.php)
- [Concrete wall drift: American Society of Civil Engineers (ASCE) 7-22—0.005h, Model?](https://aistructuralreview.com/blog/concrete-wall-drift-american-society-of-civil-engineers-asce-7-220005h-model.php)

### Latest

- [Stronger Thigh Bones: Evolutionary Structural Optimization (ESO) — 2% Removal...](https://aistructuralreview.com/blog/stronger-thigh-bones-evolutionary-structural-optimization-eso-2-removal-or-not.php)
- [Memphis building codes: 2018 International Building Code lateral retrofit vs...](https://aistructuralreview.com/blog/memphis-building-codes-2018-international-building-code-lateral-retrofit-vs-rebuild.php)
- [2% Drift Limit: Verified Gate, Not Default, for 8-Story Steel Frames](https://aistructuralreview.com/blog/2-drift-limit-verified-gate-not-default-for-8-story-steel-frames.php)

Canonical: https://aistructuralreview.com/blog/pinn-drift-surrogates-pass-a-20-held-out-midr-test-first.php
Markdown: https://aistructuralreview.com/blog/pinn-drift-surrogates-pass-a-20-held-out-midr-test-first.php/index.md